I come to terms that it's maybe not impossible, but most certainly too much hassle to make something work like i want it too.
After a lot of custom scripts, trying different approaches, text encoder training , changing how trainers actually train models, to a point where i made a lora that should not work but it did...
Training just a few images that have a constant theme, like photorealistic, is no problem. A big dataset which maybe has the same theme, but the variation in style are so extreme, require a lot of steps and care with the learning rate, resolutions, sometimes captations are to "accurate" which can lead to early overfitting ( streaks, artifacts ) not enough time to learn the picture before collapse, yadayada.
Baseline is, Chroma still suffers from the model it was trained on, even now. Low Steps, Destillation ( Flux Schnell ) and maybe even locked weights are absolute poison for training, because that stuff restricts, makes a model rigged and narrows down the variations. Also a reason why i stay away from Z-Image for now till they release the base. But even then it's Alibaba with the good ol' chinese censorship and tons of concepts it doesn't even know, which most likely requires a few months of fine tuning to get a real "Base". Really hope that happens, because it's more light-weight, good quality and has a very good understanding of prompts. Who knows.
Maybe more my own fault then the model itself, trying to brute force something that in general is a bad approach to training... just a reason to do it anyway :)
Description
FAQ
Comments (5)
👍🏻👍🏻👍🏻
can you make a chroma lora that just focus dark light, night light, night setting.
Chroma or Flux in generell has no problem with dark or night scenes, depends on how the prompt looks. But, i'm testing slider loras at the moment. If i get the time i will try to sneak a slider for night, day - dark, bright etc in... maybe not the worst idea.
@TijuanaSlumlord it difficult to fight against the constant spotlight that shines directly on the subject regardless of hard i prompt for a very dark night setting image.
@obinna7713 Just asking for a night scene is in a lot of cases not enough, doesn't matter if it's a model based on SDXL, Flux or anything else. If you look trough some of the prompts of the images i posted, you may notice how often different light sources are mentioned and what type of light sources.
If you just ask for "night" or "dark" those are extreme global and broad light conditions, so the model takes what it knows and adds the light source however it wants. A lot of images that have been trained in models have this "spotlight" or "flash photography" style, which is just one bright overpowering light source that produces hard shadows and makes a character or object stick out while killing everything else in the background or surrounding.
You don't have to go as crazy with it as in my prompts, maybe just try to add a light source or multiple light sources that produce soft shadows, like a candle or torch, something that glows, ethereal light, a dim streetlight, bathed in cool blue ( or whatever color ) tones, so basically light that spreads across the whole scene or something dim that produces soft shadows and not sterile studio or spotlight.